融合多元异构信息的矩阵分解推荐算法  

Matrix Factorization Recommendation Algorithms Based on Multiple Heterogeneous Information Fusion

作　　者：王根生 潘方正[1] WANG Gen-sheng;PAN Fang-zheng(School of Humanities,Jiangxi University of Finance and Economic,Nanchang 330013,China;Computer Practice Teaching Center,Jiangxi University of Finance and Economics,Nanchang 330013,China;School of International Trade and Economics,Jiangxi University of Finance and Economic,Nanchang 330013,China)

机构地区：[1]江西财经大学人文学院,南昌330013 [2]江西财经大学计算机实践教学中心,南昌330013 [3]江西财经大学国际经贸学院,南昌330013

出　　处：《小型微型计算机系统》2020年第7期1406-1412,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(71461012)资助;江西省高校人文社会科学研究项目(GL19110)资助。

摘　　要：针对矩阵分解推荐算法存在的数据稀疏和不能反映用户兴趣变化的问题,提出一种融合用户点评数据、用户-物品评分数据、物品异构信息和遗忘曲线的改进型矩阵分解推荐算法.首先,利用Worde2Vec对用户点评数据中包含的物品内容信息进行挖掘,结合物品属性数据得出物品相似度矩阵;然后,构建用户-物品-属性的加权异构信息网络,在权重计算时引入激活函数和遗忘曲线,通过元路径计算用户相似度矩阵;最后,把物品相似度矩阵和用户相似度矩阵融合到矩阵分解算法的目标函数中.实验结果表明,该改进算法相比于传统矩阵分解推荐算法和部分其他改进算法具有更高的准确率、召回率、覆盖率和更低的均方根误差.In order to solve the problem of sparse data and not considering the drifting of user’s interest in the matrix factorization recommendation algorithm,an improved matrix factorization recommendation algorithm based on user’s comment data,user’s score data,heterogeneous information of goods and forgetting curve is proposed.Firstly,using Worde2vec to mine the content information of the items contained in the user’s comment data,and combining the item attribute data to get the item similarity matrix;Secondly,using the activation function and forgetting curve to calculate the weight to construct the user-item-attribute weighted heterogeneous information network,and calculate the user similarity matrix through meta path;Finally,integrating the similarity matrix of items and users into the objective function of matrix factorization algorithm.The experimental results show that the improved algorithm has higher accuracy,recall and coverage and lower root mean square error than the traditional matrix decomposition recommendation algorithm and some other improved algorithms.

关 键 词：推荐算法 矩阵分解 异构信息 兴趣变化 遗忘曲线 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]